Home
Purpose of Conference
Topics
Committees
CISIM Dates
Final Version Submission
Accepted Papers
Registration and Fees
PROGRAMME
Keynote Speakers
Proceedings
Postconference Publication
Venue
Access
Accommodation
Previous Conferences
PDF version of Call for Papers

Krzysztof Cios is Professor and Chair of Computer Science Department at the Virginia Commonwealth University, Richmond, U.S.A. His research is in the areas of machine learning, data mining, computational neuroscience, and biomedical informatics, and it was funded by NIH, NASA, NSF, NATO, and the U.S. Air Force. His former students work as professors at the American, Australian and Thai universities, as post-doctoral researchers at Harvard, Johns Hopkins and NIH, or at companies such as NASA, General Motors and Proctor and Gamble. He published extensively and has been the recipient of the Norbert Wiener Outstanding Paper Award, the Neurocomputing Best Paper Award, and the Fulbright Senior Scholar Award. He is a Foreign Member of the Polish Academy of Arts and Sciences (PAU).

Building Data Models with Rule Learners: Classical, Multiple-Instance, and One-Class Learning Algorithms

First, we shall talk about supervised inductive machine learning algorithms that generate rules and explain why rule learners are a preferred choice for model building in domains where understanding of a model is important, such as in medicine. Then we will introduce a classical rule learner that is scalable to big data. Note that classical rule learners require knowledge about class memberships of all instances. Next, we will introduce challenging multiple-instance learning (MIL) and one-class learning problems. The MIL is concerned with classifying bags of instances instead of single instances. A bag is labeled as positive if at least one of its instances is positive, and as negative if all of its instances are negative. In a one-class scenario only a single (target) class of instances is available; this type of learning is also known as an outlier, or novelty, detection problem. Since most inductive machine learning algorithms require discretization as a pre-processing step we will briefly describe an information-theoretic algorithm that uses class information to automatically generate a number of intervals for a given attribute. Second, we shall present MIL and one-class algorithms and introduce a general framework for converting classical algorithms into such algorithms.